Contact us
How we build and deliver
Learn more about us and get an architect-vetted startup founder checklist
This website uses cookies to ensure you get the best experience
AI governance consulting and implementation
Set up an inventory of your AI systems, define evaluation and review processes, and add monitoring. We implement technical controls and document responsibilities across your organization.
The core engineering challenge
Deploying machine learning models in production introduces non-deterministic risks that traditional software compliance cannot catch. You need more than a checklist; you need a technical architecture that enforces safety.
We implement technical controls in your AI delivery and monitoring processes. The goal is to make risks, model changes and exceptions visible, with documented responsibilities and review steps.
Technical stack and instrumentation
Algorithmic auditing and fairness
We test for disparate impact and bias across protected groups
- IBM AI Fairness 360 (AIF360): We use this to detect and fix bias in datasets and models.
- Fairlearn: We apply this for group fairness metrics assessment during model selection.
- What-If Tool (WIT): We use this for probing model behavior across different hypothetical situations.
SHAP (SHapley Additive exPlanations): We calculate the contribution of each feature to the prediction.
LIME (Local Interpretable Model-agnostic Explanations): We use this to approximate the model locally and explain individual predictions.
LIME (Local Interpretable Model-agnostic Explanations): We use this to approximate the model locally and explain individual predictions.
- ELI5: We deploy this to debug machine learning classifiers and check their inference steps.
Explainability and interpretability (XAI)
We make "black box" models transparent so stakeholders understand why a decision was made
Data privacy and security
We secure the data lineage (the lifecycle of data origin and movement) and prevent leakage
- TensorFlow Privacy: We apply differential privacy (adding noise to obscure individual data points) to train models without exposing user data.
- PySyft: We use this for encrypted, privacy-preserving deep learning.
- CleverHans: We test your models against adversarial examples (inputs designed to trick the model) to ensure robustness.
Our execution workflow
We use a four-phase process to audit, fix, and maintain your AI infrastructure.
Discovery and taxonomy
Inventory AI systems, their intended uses, data sources and owners. We document relevant technical risks and help map evidence needs for your organization’s review process.
Stress test
We try to break your models.
Bias Testing: We run your models against synthetic datasets to check for discrimination.
Adversarial Attack Simulation: We inject noise and edge cases to see if the model fails.
Code Review: We analyze your Jupyter notebooks and training scripts for reproducibility and security flaws.
Bias Testing: We run your models against synthetic datasets to check for discrimination.
Adversarial Attack Simulation: We inject noise and edge cases to see if the model fails.
Code Review: We analyze your Jupyter notebooks and training scripts for reproducibility and security flaws.
Remediation and hardening
Prioritize the findings, evaluate potential model or data changes, document limitations and configure appropriate access controls.
Continuous monitoring
Monitor agreed performance and risk indicators, route exceptions for review and record relevant system and version changes.
Frequently asked questions
AI safety work evaluates potential harms and failure modes. Compliance work addresses applicable obligations and organizational requirements. We provide technical assessments and implementation support for these reviews.
The scope can include an AI system inventory, named owners, evaluation criteria, review procedures and monitoring. We agree which technical controls and evidence your organization needs before implementation.
We can evaluate outputs, test representative cases and use suitable explainability tools. These methods have limits, which we document alongside the findings.
You should audit continuously. One-time audits fail because data changes. We recommend setting up automated monitoring that runs daily, with a deep manual audit performed quarterly or whenever you push a major version update to production.
No. We provide technical engineering and assessment services. We work with your legal and compliance teams on the controls, documentation and evidence they identify as necessary.
What needs more visibility or control?
Tell us about your AI systems, current monitoring and the review you are preparing for. We’ll discuss the technical scope and useful evidence.